改进Faster RCNN with FPN的素布瑕疵检测的算法研究

Research on algorithm for improving Faster RCNN with FPN for plain cloth defect detection

纺织行业中的布匹检测仍存在采用人工检测的情况,人工检测效果受工人主观影响较大,易发生检测效率的降低和瑕疵的漏检误检。针对这种现状,探究素布瑕疵检测的算法,改进Faster RCNNwith FPN目标检测算法。首先,为了提升Faster RCNNwithFPN对于多尺度特征的融合能力,丰富各个特征层的上下文信息,引入跨尺度特征融合模块来改进特征金字塔网络结构。其次,为了更好的利用深层特征,加入尺度内特征交互模块来处理ResNet50输出的深层特征层,丰富高级特征层的语义信息。然后,为了增强对于极端尺寸瑕疵目标的检测能力,使用K-means++聚类和遗传算法,改进预设锚框。最后,由于素布瑕疵的尺寸较小,为了平衡正负样本,采用Focal Loss,增加对于素布瑕疵的检测效果。经过实验,使用COCO指标进行评价,该改进后的网络模型与Faster RCNNwithFPN相比,在mAP_(50)、mAP_(75)和mAP_(50:95)指标上分别提升6.5%、4.4%和4.0%,平均准确率有了明显提升,可以更好地完成素布瑕疵的检测任务。

In the textile industry,manual detection is still used for cloth inspection.The effect of manual detec-tion is greatly affected by workers'subjectivity,leading to reduced efficiency and missed or erroneous detection of defects.To address this situation,the algorithm of plain cloth defect detection is explored to improve Faster RCNN with FPN target detection algorithm.Firstly,in order to improve the fusion capability of Faster RCNN with FPN for multi-scale features and enrich the context information of each feature layer,the cross-scale fea-ture fusion module is introduced to improve the feature pyramid network structure.Secondly,in order to make better use of deep features,the intra-scale feature interaction module is added to process the deep feature layer output by ResNet50,and enrich the semantic information of the high-level feature layer.Then,in order to en-hance the detection capability for defects of extreme sizes,preset anchor boxes are improved by using K means++clustering and genetic algorithms.Finally,considering the small size of plain fabric defects,Focal Loss is used to balance the positive and negative samples to increase the detection effect of plain cloth defects.Through experiments,the COCO index is used for evaluation.Compared with Faster RCNN with FPN,the im-proved network model increases by 6.5%,4.4%and 4.0%on the mAP_(50),mAP_(75)and mAP_(50:95)indexes,respective-ly.The average accuracy has been significantly improved,which can better complete the detection task of plain cloth defects.